Most AI products hide their economics. OnyxBot shows you the exact price of every AI call it makes for your server, down to fractions of a cent, and what each answer was built from.
Where to look
- Home, then Usage is the ledger. Every AI call is a row: when it ran, what it did, the feature it belongs to, the model, the tokens, the cost, and whether it was billed.
- Billing, in your account menu, shows live usage against your included amount and your cap, including exact overage so far. It is always the full picture.
Reading the Usage page
The period switcher at the top offers 24h, 7 days, 30 days, and This period, your current billing cycle. Four tiles sum the period up: Spend this period, the change vs previous period, the Top cost driver, and how much was Not billed because it fell inside your included usage or OnyxBot covered it. Where it went splits the spend by feature: chat replies in public channels, tickets, knowledge, insights, and GitHub. Click a segment to filter the ledger to that feature. The chart under it shows spend by day.
The ledger takes filters for feature, billing status, and model, and sorts by newest or by cost. Load more pages further back, and the download button exports the filtered rows as CSV. Per-call records begin on the day recording started for your server. Spend before that day is in your Billing total but is not itemized here.
Click any row for its detail panel: what the activity cost and whether it was billed, every call that made it up with its model and tokens, and a link to the ticket transcript when it ran inside a ticket. For a chat or ticket answer the panel also lists its Sources: what the search retrieved, what made it into the prompt, and what the model cited, with a Show the question and answer toggle. That is how you find the page behind a wrong answer. The block at the bottom has the model, the provider, the token counts with how many were read from the cache, the time to answer, and a reference id to quote to support.
What drives the cost
Longer questions, longer answers, and stronger models cost more. A typical grounded support answer on the default model costs well under a cent. Filtered messages (greetings and chatter the AI declines) cost nothing.
An answer is usually several calls: the screening step, the knowledge search, the reply itself, and a summary afterwards. The ledger groups them under one row and the detail panel lists each. Background jobs are billed too, and each has its own row: the staff briefing written when a ticket escalates, the lesson pulled from a resolved ticket, the summary that seeds a ticket opened from a public channel, the GitHub plugin's digests, and the Insights reports, if you use them. Filter by feature to see any of them on their own.
Why we do this
Metered billing only feels fair when you can audit it. Seeing per-answer costs also makes decisions easy: whether a premium model is worth it, whether your included usage fits your volume, and what a busy month actually costs you.